enum radeon_bo_usage rusage = RADEON_USAGE_READWRITE;
bool busy = false;
+ assert(!(resource->flags & RADEON_FLAG_SPARSE));
+
if (usage & PIPE_TRANSFER_UNSYNCHRONIZED) {
return ctx->ws->buffer_map(resource->buf, NULL, usage);
}
res->bo_size = size;
res->bo_alignment = alignment;
res->flags = 0;
+ res->texture_handle_allocated = false;
+ res->image_handle_allocated = false;
switch (res->b.b.usage) {
case PIPE_USAGE_STREAM:
res->flags |= RADEON_FLAG_GTT_WC;
break;
}
- res->flags |= RADEON_FLAG_CPU_ACCESS;
/* fall through */
case PIPE_USAGE_DEFAULT:
case PIPE_USAGE_IMMUTABLE:
if (rscreen->info.drm_major == 2 &&
rscreen->info.drm_minor < 40)
res->domains = RADEON_DOMAIN_GTT;
- else if (res->domains & RADEON_DOMAIN_VRAM)
- res->flags |= RADEON_FLAG_CPU_ACCESS;
}
/* Tiled textures are unmappable. Always put them in VRAM. */
- if (res->b.b.target != PIPE_BUFFER &&
- rtex->surface.level[0].mode >= RADEON_SURF_MODE_1D) {
+ if ((res->b.b.target != PIPE_BUFFER && !rtex->surface.is_linear) ||
+ res->flags & R600_RESOURCE_FLAG_UNMAPPABLE) {
res->domains = RADEON_DOMAIN_VRAM;
- res->flags &= ~RADEON_FLAG_CPU_ACCESS;
res->flags |= RADEON_FLAG_NO_CPU_ACCESS |
RADEON_FLAG_GTT_WC;
}
/* If VRAM is just stolen system memory, allow both VRAM and
* GTT, whichever has free space. If a buffer is evicted from
* VRAM to GTT, it will stay there.
+ *
+ * DRM 3.6.0 has good BO move throttling, so we can allow VRAM-only
+ * placements even with a low amount of stolen VRAM.
*/
if (!rscreen->info.has_dedicated_vram &&
- res->domains == RADEON_DOMAIN_VRAM)
+ (rscreen->info.drm_major < 3 || rscreen->info.drm_minor < 6) &&
+ res->domains == RADEON_DOMAIN_VRAM) {
res->domains = RADEON_DOMAIN_VRAM_GTT;
+ res->flags &= ~RADEON_FLAG_NO_CPU_ACCESS; /* disallowed with VRAM_GTT */
+ }
if (rscreen->debug_flags & DBG_NO_WC)
res->flags &= ~RADEON_FLAG_GTT_WC;
+ if (res->b.b.bind & PIPE_BIND_SHARED)
+ res->flags |= RADEON_FLAG_NO_SUBALLOC;
+
/* Set expected VRAM and GART usage for the buffer. */
res->vram_usage = 0;
res->gart_usage = 0;
{
struct r600_resource *rbuffer = r600_resource(buf);
+ threaded_resource_deinit(buf);
util_range_destroy(&rbuffer->valid_buffer_range);
pb_reference(&rbuffer->buf, NULL);
FREE(rbuffer);
struct r600_resource *rbuffer)
{
/* Shared buffers can't be reallocated. */
- if (rbuffer->is_shared)
+ if (rbuffer->b.is_shared)
+ return false;
+
+ /* Sparse buffers can't be reallocated. */
+ if (rbuffer->flags & RADEON_FLAG_SPARSE)
return false;
/* In AMD_pinned_memory, the user pointer association only gets
* broken when the buffer is explicitly re-allocated.
*/
- if (rctx->ws->buffer_is_user_ptr(rbuffer->buf))
+ if (rbuffer->b.is_user_ptr)
return false;
/* Check if mapping this buffer would cause waiting for the GPU. */
return true;
}
+/* Replace the storage of dst with src. */
+void r600_replace_buffer_storage(struct pipe_context *ctx,
+ struct pipe_resource *dst,
+ struct pipe_resource *src)
+{
+ struct r600_common_context *rctx = (struct r600_common_context *)ctx;
+ struct r600_resource *rdst = r600_resource(dst);
+ struct r600_resource *rsrc = r600_resource(src);
+ uint64_t old_gpu_address = rdst->gpu_address;
+
+ pb_reference(&rdst->buf, rsrc->buf);
+ rdst->gpu_address = rsrc->gpu_address;
+
+ assert(rdst->vram_usage == rsrc->vram_usage);
+ assert(rdst->gart_usage == rsrc->gart_usage);
+ assert(rdst->bo_size == rsrc->bo_size);
+ assert(rdst->bo_alignment == rsrc->bo_alignment);
+ assert(rdst->domains == rsrc->domains);
+ assert(rdst->flags == rsrc->flags);
+
+ rctx->rebind_buffer(ctx, dst, old_gpu_address);
+}
+
void r600_invalidate_resource(struct pipe_context *ctx,
struct pipe_resource *resource)
{
static void *r600_buffer_get_transfer(struct pipe_context *ctx,
struct pipe_resource *resource,
- unsigned level,
unsigned usage,
const struct pipe_box *box,
struct pipe_transfer **ptransfer,
unsigned offset)
{
struct r600_common_context *rctx = (struct r600_common_context*)ctx;
- struct r600_transfer *transfer = slab_alloc_st(&rctx->pool_transfers);
-
- transfer->transfer.resource = resource;
- transfer->transfer.level = level;
- transfer->transfer.usage = usage;
- transfer->transfer.box = *box;
- transfer->transfer.stride = 0;
- transfer->transfer.layer_stride = 0;
+ struct r600_transfer *transfer;
+
+ if (usage & TC_TRANSFER_MAP_THREADED_UNSYNC)
+ transfer = slab_alloc(&rctx->pool_transfers_unsync);
+ else
+ transfer = slab_alloc(&rctx->pool_transfers);
+
+ transfer->b.b.resource = NULL;
+ pipe_resource_reference(&transfer->b.b.resource, resource);
+ transfer->b.b.level = 0;
+ transfer->b.b.usage = usage;
+ transfer->b.b.box = *box;
+ transfer->b.b.stride = 0;
+ transfer->b.b.layer_stride = 0;
+ transfer->b.staging = NULL;
transfer->offset = offset;
transfer->staging = staging;
- *ptransfer = &transfer->transfer;
+ *ptransfer = &transfer->b.b;
return data;
}
{
struct r600_common_context *rctx = (struct r600_common_context*)ctx;
struct r600_common_screen *rscreen = (struct r600_common_screen*)ctx->screen;
- struct r600_resource *rbuffer = r600_resource(resource);
- uint8_t *data;
+ struct r600_resource *rbuffer = r600_resource(resource);
+ uint8_t *data;
assert(box->x + box->width <= resource->width0);
+ /* From GL_AMD_pinned_memory issues:
+ *
+ * 4) Is glMapBuffer on a shared buffer guaranteed to return the
+ * same system address which was specified at creation time?
+ *
+ * RESOLVED: NO. The GL implementation might return a different
+ * virtual mapping of that memory, although the same physical
+ * page will be used.
+ *
+ * So don't ever use staging buffers.
+ */
+ if (rbuffer->b.is_user_ptr)
+ usage |= PIPE_TRANSFER_PERSISTENT;
+
/* See if the buffer range being mapped has never been initialized,
* in which case it can be mapped unsynchronized. */
- if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
+ if (!(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
+ TC_TRANSFER_MAP_IGNORE_VALID_RANGE)) &&
usage & PIPE_TRANSFER_WRITE &&
- !rbuffer->is_shared &&
+ !rbuffer->b.is_shared &&
!util_ranges_intersect(&rbuffer->valid_buffer_range, box->x, box->x + box->width)) {
usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
}
}
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE &&
- !(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
+ !(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
+ TC_TRANSFER_MAP_NO_INVALIDATE))) {
assert(usage & PIPE_TRANSFER_WRITE);
if (r600_invalidate_buffer(rctx, rbuffer)) {
}
if ((usage & PIPE_TRANSFER_DISCARD_RANGE) &&
- !(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
- PIPE_TRANSFER_PERSISTENT)) &&
!(rscreen->debug_flags & DBG_NO_DISCARD_RANGE) &&
- r600_can_dma_copy_buffer(rctx, box->x, 0, box->width)) {
+ ((!(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
+ PIPE_TRANSFER_PERSISTENT)) &&
+ r600_can_dma_copy_buffer(rctx, box->x, 0, box->width)) ||
+ (rbuffer->flags & RADEON_FLAG_SPARSE))) {
assert(usage & PIPE_TRANSFER_WRITE);
- /* Check if mapping this buffer would cause waiting for the GPU. */
- if (r600_rings_is_buffer_referenced(rctx, rbuffer->buf, RADEON_USAGE_READWRITE) ||
+ /* Check if mapping this buffer would cause waiting for the GPU.
+ */
+ if (rbuffer->flags & RADEON_FLAG_SPARSE ||
+ r600_rings_is_buffer_referenced(rctx, rbuffer->buf, RADEON_USAGE_READWRITE) ||
!rctx->ws->buffer_wait(rbuffer->buf, 0, RADEON_USAGE_READWRITE)) {
/* Do a wait-free write-only transfer using a temporary buffer. */
unsigned offset;
struct r600_resource *staging = NULL;
- u_upload_alloc(rctx->uploader, 0, box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT),
- 256, &offset, (struct pipe_resource**)&staging, (void**)&data);
+ u_upload_alloc(ctx->stream_uploader, 0,
+ box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT),
+ rctx->screen->info.tcc_cache_line_size,
+ &offset, (struct pipe_resource**)&staging,
+ (void**)&data);
if (staging) {
data += box->x % R600_MAP_BUFFER_ALIGNMENT;
- return r600_buffer_get_transfer(ctx, resource, level, usage, box,
+ return r600_buffer_get_transfer(ctx, resource, usage, box,
ptransfer, data, staging, offset);
+ } else if (rbuffer->flags & RADEON_FLAG_SPARSE) {
+ return NULL;
}
} else {
/* At this point, the buffer is always idle (we checked it above). */
usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
}
}
- /* Using a staging buffer in GTT for larger reads is much faster. */
- else if ((usage & PIPE_TRANSFER_READ) &&
- !(usage & (PIPE_TRANSFER_WRITE |
- PIPE_TRANSFER_PERSISTENT)) &&
- rbuffer->domains & RADEON_DOMAIN_VRAM &&
- r600_can_dma_copy_buffer(rctx, 0, box->x, box->width)) {
+ /* Use a staging buffer in cached GTT for reads. */
+ else if (((usage & PIPE_TRANSFER_READ) &&
+ !(usage & PIPE_TRANSFER_PERSISTENT) &&
+ (rbuffer->domains & RADEON_DOMAIN_VRAM ||
+ rbuffer->flags & RADEON_FLAG_GTT_WC) &&
+ r600_can_dma_copy_buffer(rctx, 0, box->x, box->width)) ||
+ (rbuffer->flags & RADEON_FLAG_SPARSE)) {
struct r600_resource *staging;
+ assert(!(usage & TC_TRANSFER_MAP_THREADED_UNSYNC));
staging = (struct r600_resource*) pipe_buffer_create(
ctx->screen, 0, PIPE_USAGE_STAGING,
box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT));
if (staging) {
/* Copy the VRAM buffer to the staging buffer. */
- ctx->resource_copy_region(ctx, &staging->b.b, 0,
- box->x % R600_MAP_BUFFER_ALIGNMENT,
- 0, 0, resource, level, box);
+ rctx->dma_copy(ctx, &staging->b.b, 0,
+ box->x % R600_MAP_BUFFER_ALIGNMENT,
+ 0, 0, resource, 0, box);
- data = r600_buffer_map_sync_with_rings(rctx, staging, PIPE_TRANSFER_READ);
+ data = r600_buffer_map_sync_with_rings(rctx, staging,
+ usage & ~PIPE_TRANSFER_UNSYNCHRONIZED);
if (!data) {
r600_resource_reference(&staging, NULL);
return NULL;
}
data += box->x % R600_MAP_BUFFER_ALIGNMENT;
- return r600_buffer_get_transfer(ctx, resource, level, usage, box,
+ return r600_buffer_get_transfer(ctx, resource, usage, box,
ptransfer, data, staging, 0);
+ } else if (rbuffer->flags & RADEON_FLAG_SPARSE) {
+ return NULL;
}
}
}
data += box->x;
- return r600_buffer_get_transfer(ctx, resource, level, usage, box,
+ return r600_buffer_get_transfer(ctx, resource, usage, box,
ptransfer, data, NULL, 0);
}
struct pipe_transfer *transfer,
const struct pipe_box *rel_box)
{
- if (transfer->usage & (PIPE_TRANSFER_WRITE |
- PIPE_TRANSFER_FLUSH_EXPLICIT)) {
+ unsigned required_usage = PIPE_TRANSFER_WRITE |
+ PIPE_TRANSFER_FLUSH_EXPLICIT;
+
+ if ((transfer->usage & required_usage) == required_usage) {
struct pipe_box box;
u_box_1d(transfer->box.x + rel_box->x, rel_box->width, &box);
!(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT))
r600_buffer_do_flush_region(ctx, transfer, &transfer->box);
- if (rtransfer->staging)
- r600_resource_reference(&rtransfer->staging, NULL);
+ r600_resource_reference(&rtransfer->staging, NULL);
+ assert(rtransfer->b.staging == NULL); /* for threaded context only */
+ pipe_resource_reference(&transfer->resource, NULL);
- slab_free_st(&rctx->pool_transfers, transfer);
+ /* Don't use pool_transfers_unsync. We are always in the driver
+ * thread. */
+ slab_free(&rctx->pool_transfers, transfer);
}
void r600_buffer_subdata(struct pipe_context *ctx,
rbuffer->b.b.next = NULL;
pipe_reference_init(&rbuffer->b.b.reference, 1);
rbuffer->b.b.screen = screen;
+
rbuffer->b.vtbl = &r600_buffer_vtbl;
+ threaded_resource_init(&rbuffer->b.b);
+
rbuffer->buf = NULL;
rbuffer->bind_history = 0;
rbuffer->TC_L2_dirty = false;
- rbuffer->is_shared = false;
util_range_init(&rbuffer->valid_buffer_range);
return rbuffer;
}
r600_init_resource_fields(rscreen, rbuffer, templ->width0, alignment);
- if (templ->bind & PIPE_BIND_SHARED)
- rbuffer->flags |= RADEON_FLAG_HANDLE;
+ if (templ->flags & PIPE_RESOURCE_FLAG_SPARSE)
+ rbuffer->flags |= RADEON_FLAG_SPARSE;
if (!r600_alloc_resource(rscreen, rbuffer)) {
FREE(rbuffer);
}
struct pipe_resource *r600_aligned_buffer_create(struct pipe_screen *screen,
- unsigned bind,
+ unsigned flags,
unsigned usage,
unsigned size,
unsigned alignment)
memset(&buffer, 0, sizeof buffer);
buffer.target = PIPE_BUFFER;
buffer.format = PIPE_FORMAT_R8_UNORM;
- buffer.bind = bind;
+ buffer.bind = 0;
buffer.usage = usage;
- buffer.flags = 0;
+ buffer.flags = flags;
buffer.width0 = size;
buffer.height0 = 1;
buffer.depth0 = 1;
struct r600_resource *rbuffer = r600_alloc_buffer_struct(screen, templ);
rbuffer->domains = RADEON_DOMAIN_GTT;
+ rbuffer->flags = 0;
+ rbuffer->b.is_user_ptr = true;
util_range_add(&rbuffer->valid_buffer_range, 0, templ->width0);
+ util_range_add(&rbuffer->b.valid_buffer_range, 0, templ->width0);
/* Convert a user pointer to a buffer. */
rbuffer->buf = ws->buffer_from_ptr(ws, user_memory, templ->width0);
else
rbuffer->gpu_address = 0;
+ rbuffer->vram_usage = 0;
+ rbuffer->gart_usage = templ->width0;
+
return &rbuffer->b.b;
}
projects / mesa.git / blobdiff